Slashdot Mirror


Weta Prepares to Render LOTR: ROTK

Dee Arsmith writes "Peter Jackson's special-effects company Weta Digital has just taken delivery of 588 IBM blade servers, each with two 2.8 gigahertz Intel Xeon processors. Seven racks of IBM blade servers have been added to Weta's existing 15-rack server cluster to make up the largest Intel-based high- performance computer site in the world with more than 2000 linked processors. The cluster will be used to render the frames drawn by the animators to complete the final installment of The Lord of the Rings trilogy, The Return of the King."

4 of 460 comments (clear)

  1. Really the top? by cly · · Score: 5, Informative

    If you look at top500.org, you see that the current top Intel-based cluster is #5, the one with 2304 procs in LLNL.

    The article says their cluster has 'more than 2000 processors'. So presumably they mean 'more than 2304'?

  2. Re:Question by zokrath · · Score: 5, Informative

    Bah, last post was apparently in HTML format and managed to make a single unreadable block of text. That will teach me not to preview...Anyway, reposted in plain text:

    A long, long ways.

    Computer games can run at 60+ frames per second because they are barely doing any work when compared to top of the line rendering engines.

    Raytracing, dozens of texture passes, multiple realistic lightsources; and these are just for a two dimensional surface. Making realistic looking skin requires multiple translucent layers to simulate the complicated appearance of skin.

    Also, there is the size factor. Video games generally run at 1024x768 to 1600x1200. Movie quality shots are rendered at many times that resolution, which greatly increases the number of pixels that have to be rendered. Gollum may only be 800 pixels tall on your monitor, but he's probably rendered at least ten times as large; we'll say 10,000x 10,000 for calcualtion simplification.

    That's 10E7 pixels, so to display it at 24 frames per second you would need to be pushing 24E8 pixels a second. 24,000,000,000.

    Even if every pixel only took a single cycle (which it might, with the right hardware pipeline in the future), you would need 240 terahertz of power (plus overhead) to display it in real time, along with enough RAM to hold the model and texture data for everything that's going to be onscreen within the next minute or so.

    Considering that they have around 2000 x 2.0 X 2 = 8 terahertz available to them, and it still takes ages to render each frame of the complicated battle scenes, I'd say we are going to hit the limit of Moore's law before we could reasonably get hte power to render cinematic scenes in real time. Perhaps with quantum processing we will be able to within the next 20 years or so.

  3. Re:588??? by Tihstae · · Score: 5, Informative

    I would just like to know why 588 computers?

    IBM Blade Center that holds the blade server is 7U. Each Blade Center holds 14 blade servers. IBM's racks are 42U.

    42U Rack / 7U Blade Center = 6 Blade Centers/rack
    14 servers X 6 Blade Centers = 84 servers/rack
    7 Racks X 84 servers = 588 Servers

  4. Re:Question by donglekey · · Score: 5, Informative

    Also, there is the size factor. Video games generally run at 1024x768 to 1600x1200. Movie quality shots are rendered at many times that resolution, which greatly increases the number of pixels that have to be rendered. Gollum may only be 800 pixels tall on your monitor, but he's probably rendered at least ten times as large; we'll say 10,000x 10,000 for calcualtion simplification.

    This is actually not true. Film resolution is around 2048x1556 and everything is rendered the size that it is needed. For the most part, the difference in rendering speed is because hardware is very fast and very efficient, and takes lots of shortcuts. There aren't many textures, they aren't very high resolution, there isn't any raytracing, there are very few lights, no global illumination, no hair rendering, no volumetric rendering, not nearly as many polygons, no particles or cloth simulations, very few deformations, and lighting calculation is done on vertexes and then interpolated instead of on every pixel (this will change with Doom 3 and Half Life 2 which is the real reason they look so much better). Renderman also subdivides everything down to one polygon per pixel to get perfectly smooth sufaces and good displacement. There is also the issue of motion blur, depth of field, and rendering of composites, which also takes a very long time. Anti-aliasing in every step is crucial for any kind of non-realtime CG, but it not as important for games, and that by itself makes a huge huge difference.